← All resources

Real-time Rendering with a Neural Irradiance Volume

Auto-discovered

Paper Advanced

Arno Coomans, Giacomo Nazzaro, Edoardo A. Dominici, Christian Döring, Floor Verhoeven, Konstantinos Vardis, Markus Steinberger

Visit resource ↗

Description

Real-time Rendering with a Neural Irradiance Volume is the Eurographics 2026 paper by Arno Coomans, Giacomo Nazzaro, and coauthors from Huawei Technologies and Graz University of Technology presenting Neural Irradiance Volume (NIV) for diffuse global illumination.

NIV replaces irradiance probe grids with a compact neural model that precomputes and compresses scene irradiance, including dynamic objects. The abstract reports about 1 ms per frame on consumer GPUs at full HD, roughly 1-5 MB for medium scenes, G-buffer-only input without ray tracing or denoising, and large quality gains versus probes at equal memory.

It targets advanced real-time rendering researchers and engine developers under strict frame-time and memory budgets. Familiarity with probe volumes, deferred G-buffers, and neural compression helps.

The project page provides the abstract, TL;DR, video abstract, and an arXiv link for the full paper.

The work is a concrete neural alternative to cubic probe grids when memory scaling and aliasing dominate GI cost. Link last verified accessible on 2026-09-27. Link last verified accessible on 2026-09-28.